未验证 提交 15ce2c1b 编写于 作者: Z zhiboniu 提交者: GitHub

transfer op multiclass_nms3 to phi (#44765)

* add cmake enforce

* transfer multiclass_nms3  to phi
上级 02414aac
...@@ -123,9 +123,5 @@ cc_test( ...@@ -123,9 +123,5 @@ cc_test(
mask_util_test mask_util_test
SRCS mask_util_test.cc SRCS mask_util_test.cc
DEPS memory mask_util) DEPS memory mask_util)
cc_library(
gpc
SRCS gpc.cc
DEPS op_registry)
detection_library(generate_mask_labels_op SRCS generate_mask_labels_op.cc DEPS detection_library(generate_mask_labels_op SRCS generate_mask_labels_op.cc DEPS
mask_util) mask_util)
...@@ -13,8 +13,10 @@ limitations under the License. */ ...@@ -13,8 +13,10 @@ limitations under the License. */
#include <glog/logging.h> #include <glog/logging.h>
#include "paddle/fluid/framework/infershape_utils.h"
#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detection/nms_util.h" #include "paddle/fluid/operators/detection/nms_util.h"
#include "paddle/phi/infermeta/ternary.h"
namespace paddle { namespace paddle {
namespace operators { namespace operators {
...@@ -609,12 +611,6 @@ class MultiClassNMS3Op : public MultiClassNMS2Op { ...@@ -609,12 +611,6 @@ class MultiClassNMS3Op : public MultiClassNMS2Op {
const framework::VariableNameMap& outputs, const framework::VariableNameMap& outputs,
const framework::AttributeMap& attrs) const framework::AttributeMap& attrs)
: MultiClassNMS2Op(type, inputs, outputs, attrs) {} : MultiClassNMS2Op(type, inputs, outputs, attrs) {}
void InferShape(framework::InferShapeContext* ctx) const override {
MultiClassNMS2Op::InferShape(ctx);
ctx->SetOutputDim("NmsRoisNum", {-1});
}
}; };
class MultiClassNMS3OpMaker : public MultiClassNMS2OpMaker { class MultiClassNMS3OpMaker : public MultiClassNMS2OpMaker {
...@@ -633,6 +629,10 @@ class MultiClassNMS3OpMaker : public MultiClassNMS2OpMaker { ...@@ -633,6 +629,10 @@ class MultiClassNMS3OpMaker : public MultiClassNMS2OpMaker {
} // namespace operators } // namespace operators
} // namespace paddle } // namespace paddle
DECLARE_INFER_SHAPE_FUNCTOR(multiclass_nms3,
MultiClassNMSShapeFunctor,
PD_INFER_META(phi::MultiClassNMSInferMeta));
namespace ops = paddle::operators; namespace ops = paddle::operators;
REGISTER_OPERATOR( REGISTER_OPERATOR(
multiclass_nms, multiclass_nms,
...@@ -658,7 +658,5 @@ REGISTER_OPERATOR( ...@@ -658,7 +658,5 @@ REGISTER_OPERATOR(
ops::MultiClassNMS3Op, ops::MultiClassNMS3Op,
ops::MultiClassNMS3OpMaker, ops::MultiClassNMS3OpMaker,
paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>, paddle::framework::EmptyGradOpMaker<paddle::framework::OpDesc>,
paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>); paddle::framework::EmptyGradOpMaker<paddle::imperative::OpBase>,
REGISTER_OP_CPU_KERNEL(multiclass_nms3, MultiClassNMSShapeFunctor);
ops::MultiClassNMSKernel<float>,
ops::MultiClassNMSKernel<double>);
...@@ -21,8 +21,8 @@ limitations under the License. */ ...@@ -21,8 +21,8 @@ limitations under the License. */
namespace paddle { namespace paddle {
namespace operators { namespace operators {
using gpc::gpc_free_polygon; using phi::funcs::gpc_free_polygon;
using gpc::gpc_polygon_clip; using phi::funcs::gpc_polygon_clip;
template <class T> template <class T>
void Array2PointVec(const T* box, void Array2PointVec(const T* box,
...@@ -37,15 +37,18 @@ void Array2PointVec(const T* box, ...@@ -37,15 +37,18 @@ void Array2PointVec(const T* box,
} }
template <class T> template <class T>
void Array2Poly(const T* box, const size_t box_size, gpc::gpc_polygon* poly) { void Array2Poly(const T* box,
const size_t box_size,
phi::funcs::gpc_polygon* poly) {
size_t pts_num = box_size / 2; size_t pts_num = box_size / 2;
(*poly).num_contours = 1; (*poly).num_contours = 1;
(*poly).hole = reinterpret_cast<int*>(malloc(sizeof(int))); (*poly).hole = reinterpret_cast<int*>(malloc(sizeof(int)));
(*poly).hole[0] = 0; (*poly).hole[0] = 0;
(*poly).contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list)); (*poly).contour =
(phi::funcs::gpc_vertex_list*)malloc(sizeof(phi::funcs::gpc_vertex_list));
(*poly).contour->num_vertices = pts_num; (*poly).contour->num_vertices = pts_num;
(*poly).contour->vertex = (*poly).contour->vertex =
(gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num); (phi::funcs::gpc_vertex*)malloc(sizeof(phi::funcs::gpc_vertex) * pts_num);
for (size_t i = 0; i < pts_num; ++i) { for (size_t i = 0; i < pts_num; ++i) {
(*poly).contour->vertex[i].x = box[2 * i]; (*poly).contour->vertex[i].x = box[2 * i];
(*poly).contour->vertex[i].y = box[2 * i + 1]; (*poly).contour->vertex[i].y = box[2 * i + 1];
...@@ -53,15 +56,17 @@ void Array2Poly(const T* box, const size_t box_size, gpc::gpc_polygon* poly) { ...@@ -53,15 +56,17 @@ void Array2Poly(const T* box, const size_t box_size, gpc::gpc_polygon* poly) {
} }
template <class T> template <class T>
void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon* poly) { void PointVec2Poly(const std::vector<Point_<T>>& vec,
phi::funcs::gpc_polygon* poly) {
int pts_num = vec.size(); int pts_num = vec.size();
(*poly).num_contours = 1; (*poly).num_contours = 1;
(*poly).hole = reinterpret_cast<int*>(malloc(sizeof(int))); (*poly).hole = reinterpret_cast<int*>(malloc(sizeof(int)));
(*poly).hole[0] = 0; (*poly).hole[0] = 0;
(*poly).contour = (gpc::gpc_vertex_list*)malloc(sizeof(gpc::gpc_vertex_list)); (*poly).contour =
(phi::funcs::gpc_vertex_list*)malloc(sizeof(phi::funcs::gpc_vertex_list));
(*poly).contour->num_vertices = pts_num; (*poly).contour->num_vertices = pts_num;
(*poly).contour->vertex = (*poly).contour->vertex =
(gpc::gpc_vertex*)malloc(sizeof(gpc::gpc_vertex) * pts_num); (phi::funcs::gpc_vertex*)malloc(sizeof(phi::funcs::gpc_vertex) * pts_num);
for (size_t i = 0; i < pts_num; ++i) { for (size_t i = 0; i < pts_num; ++i) {
(*poly).contour->vertex[i].x = vec[i].x; (*poly).contour->vertex[i].x = vec[i].x;
(*poly).contour->vertex[i].y = vec[i].y; (*poly).contour->vertex[i].y = vec[i].y;
...@@ -69,7 +74,7 @@ void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon* poly) { ...@@ -69,7 +74,7 @@ void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon* poly) {
} }
template <class T> template <class T>
void Poly2PointVec(const gpc::gpc_vertex_list& contour, void Poly2PointVec(const phi::funcs::gpc_vertex_list& contour,
std::vector<Point_<T>>* vec) { std::vector<Point_<T>>* vec) {
int pts_num = contour.num_vertices; int pts_num = contour.num_vertices;
(*vec).resize(pts_num); (*vec).resize(pts_num);
...@@ -105,13 +110,13 @@ T PolyOverlapArea(const T* box1, ...@@ -105,13 +110,13 @@ T PolyOverlapArea(const T* box1,
const T* box2, const T* box2,
const size_t box_size, const size_t box_size,
const bool normalized) { const bool normalized) {
gpc::gpc_polygon poly1; phi::funcs::gpc_polygon poly1;
gpc::gpc_polygon poly2; phi::funcs::gpc_polygon poly2;
Array2Poly<T>(box1, box_size, &poly1); Array2Poly<T>(box1, box_size, &poly1);
Array2Poly<T>(box2, box_size, &poly2); Array2Poly<T>(box2, box_size, &poly2);
gpc::gpc_polygon respoly; phi::funcs::gpc_polygon respoly;
gpc::gpc_op op = gpc::GPC_INT; phi::funcs::gpc_op op = phi::funcs::GPC_INT;
gpc::gpc_polygon_clip(op, &poly2, &poly1, &respoly); phi::funcs::gpc_polygon_clip(op, &poly2, &poly1, &respoly);
T inter_area = T(0.); T inter_area = T(0.);
int contour_num = respoly.num_contours; int contour_num = respoly.num_contours;
...@@ -123,9 +128,9 @@ T PolyOverlapArea(const T* box1, ...@@ -123,9 +128,9 @@ T PolyOverlapArea(const T* box1,
inter_area += GetContourArea<T>(resvec); inter_area += GetContourArea<T>(resvec);
} }
gpc::gpc_free_polygon(&poly1); phi::funcs::gpc_free_polygon(&poly1);
gpc::gpc_free_polygon(&poly2); phi::funcs::gpc_free_polygon(&poly2);
gpc::gpc_free_polygon(&respoly); phi::funcs::gpc_free_polygon(&respoly);
return inter_area; return inter_area;
} }
......
...@@ -16,7 +16,7 @@ limitations under the License. */ ...@@ -16,7 +16,7 @@ limitations under the License. */
#include <vector> #include <vector>
#include "paddle/fluid/framework/op_registry.h" #include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/detection/gpc.h" #include "paddle/phi/kernels/funcs/gpc.h"
namespace paddle { namespace paddle {
namespace operators { namespace operators {
...@@ -47,13 +47,16 @@ void Array2PointVec(const T* box, ...@@ -47,13 +47,16 @@ void Array2PointVec(const T* box,
std::vector<Point_<T>>* vec); std::vector<Point_<T>>* vec);
template <class T> template <class T>
void Array2Poly(const T* box, const size_t box_size, gpc::gpc_polygon* poly); void Array2Poly(const T* box,
const size_t box_size,
phi::funcs::gpc_polygon* poly);
template <class T> template <class T>
void PointVec2Poly(const std::vector<Point_<T>>& vec, gpc::gpc_polygon* poly); void PointVec2Poly(const std::vector<Point_<T>>& vec,
phi::funcs::gpc_polygon* poly);
template <class T> template <class T>
void Poly2PointVec(const gpc::gpc_vertex_list& contour, void Poly2PointVec(const phi::funcs::gpc_vertex_list& contour,
std::vector<Point_<T>>* vec); std::vector<Point_<T>>* vec);
template <class T> template <class T>
......
...@@ -1670,6 +1670,15 @@ ...@@ -1670,6 +1670,15 @@
func : multi_dot func : multi_dot
backward : multi_dot_grad backward : multi_dot_grad
- api : multiclass_nms3
args : (Tensor bboxes, Tensor scores, Tensor rois_num, float score_threshold, int nms_top_k, int keep_top_k, float nms_threshold=0.3, bool normalized=true, float nms_eta=1.0, int background_label=0)
output : Tensor(out), Tensor(index), Tensor(nms_rois_num)
infer_meta :
func : MultiClassNMSInferMeta
kernel :
func : multiclass_nms3
optional : rois_num
# multinomial # multinomial
- api : multinomial - api : multinomial
args : (Tensor x, int num_samples, bool replacement) args : (Tensor x, int num_samples, bool replacement)
......
...@@ -743,6 +743,99 @@ void LinspaceInferMeta(const MetaTensor& start, ...@@ -743,6 +743,99 @@ void LinspaceInferMeta(const MetaTensor& start,
LinspaceRawInferMeta(start, stop, number, out); LinspaceRawInferMeta(start, stop, number, out);
} }
void MultiClassNMSInferMeta(const MetaTensor& bboxes,
const MetaTensor& scores,
const MetaTensor& rois_num,
float score_threshold,
int nms_top_k,
int keep_top_k,
float nms_threshold,
bool normalized,
float nms_eta,
int background_label,
MetaTensor* out,
MetaTensor* index,
MetaTensor* nms_rois_num,
MetaConfig config) {
auto box_dims = bboxes.dims();
auto score_dims = scores.dims();
auto score_size = score_dims.size();
if (config.is_runtime) {
PADDLE_ENFORCE_EQ(
score_size == 2 || score_size == 3,
true,
errors::InvalidArgument("The rank of Input(Scores) must be 2 or 3"
". But received rank = %d",
score_size));
PADDLE_ENFORCE_EQ(
box_dims.size(),
3,
errors::InvalidArgument("The rank of Input(BBoxes) must be 3"
". But received rank = %d",
box_dims.size()));
if (score_size == 3) {
PADDLE_ENFORCE_EQ(box_dims[2] == 4 || box_dims[2] == 8 ||
box_dims[2] == 16 || box_dims[2] == 24 ||
box_dims[2] == 32,
true,
errors::InvalidArgument(
"The last dimension of Input"
"(BBoxes) must be 4 or 8, "
"represents the layout of coordinate "
"[xmin, ymin, xmax, ymax] or "
"4 points: [x1, y1, x2, y2, x3, y3, x4, y4] or "
"8 points: [xi, yi] i= 1,2,...,8 or "
"12 points: [xi, yi] i= 1,2,...,12 or "
"16 points: [xi, yi] i= 1,2,...,16"));
PADDLE_ENFORCE_EQ(
box_dims[1],
score_dims[2],
errors::InvalidArgument(
"The 2nd dimension of Input(BBoxes) must be equal to "
"last dimension of Input(Scores), which represents the "
"predicted bboxes."
"But received box_dims[1](%s) != socre_dims[2](%s)",
box_dims[1],
score_dims[2]));
} else {
PADDLE_ENFORCE_EQ(box_dims[2],
4,
errors::InvalidArgument(
"The last dimension of Input"
"(BBoxes) must be 4. But received dimension = %d",
box_dims[2]));
PADDLE_ENFORCE_EQ(
box_dims[1],
score_dims[1],
errors::InvalidArgument(
"The 2nd dimension of Input"
"(BBoxes) must be equal to the 2nd dimension of Input(Scores). "
"But received box dimension = %d, score dimension = %d",
box_dims[1],
score_dims[1]));
}
}
PADDLE_ENFORCE_NE(out,
nullptr,
errors::InvalidArgument(
"The out in MultiClassNMSInferMeta can't be nullptr."));
PADDLE_ENFORCE_NE(
index,
nullptr,
errors::InvalidArgument(
"The index in MultiClassNMSInferMeta can't be nullptr."));
// Here the box_dims[0] is not the real dimension of output.
// It will be rewritten in the computing kernel.
out->set_dims(phi::make_ddim({-1, box_dims[2] + 2}));
out->set_dtype(bboxes.dtype());
index->set_dims(phi::make_ddim({-1, box_dims[2] + 2}));
index->set_dtype(DataType::INT32);
nms_rois_num->set_dims(phi::make_ddim({-1}));
nms_rois_num->set_dtype(DataType::INT32);
}
void NllLossRawInferMeta(const MetaTensor& input, void NllLossRawInferMeta(const MetaTensor& input,
const MetaTensor& label, const MetaTensor& label,
const MetaTensor& weight, const MetaTensor& weight,
......
...@@ -123,6 +123,21 @@ void LinspaceInferMeta(const MetaTensor& start, ...@@ -123,6 +123,21 @@ void LinspaceInferMeta(const MetaTensor& start,
DataType dtype, DataType dtype,
MetaTensor* out); MetaTensor* out);
void MultiClassNMSInferMeta(const MetaTensor& bboxes,
const MetaTensor& scores,
const MetaTensor& rois_num,
float score_threshold,
int nms_top_k,
int keep_top_k,
float nms_threshold,
bool normalized,
float nms_eta,
int background_label,
MetaTensor* out,
MetaTensor* index,
MetaTensor* nms_rois_num,
MetaConfig config = MetaConfig());
void NllLossRawInferMeta(const MetaTensor& input, void NllLossRawInferMeta(const MetaTensor& input,
const MetaTensor& label, const MetaTensor& label,
const MetaTensor& weight, const MetaTensor& weight,
......
...@@ -80,6 +80,7 @@ set(COMMON_KERNEL_DEPS ...@@ -80,6 +80,7 @@ set(COMMON_KERNEL_DEPS
lod_utils lod_utils
custom_kernel custom_kernel
string_infermeta string_infermeta
gpc
utf8proc) utf8proc)
copy_if_different(${kernel_declare_file} ${kernel_declare_file_final}) copy_if_different(${kernel_declare_file} ${kernel_declare_file_final})
......
此差异已折叠。
...@@ -6,6 +6,7 @@ add_subdirectory(detail) ...@@ -6,6 +6,7 @@ add_subdirectory(detail)
math_library(deformable_conv_functor DEPS dense_tensor) math_library(deformable_conv_functor DEPS dense_tensor)
math_library(concat_and_split_functor DEPS dense_tensor) math_library(concat_and_split_functor DEPS dense_tensor)
math_library(fc_functor DEPS blas jit_kernel_helper) math_library(fc_functor DEPS blas jit_kernel_helper)
math_library(gpc DEPS phi_enforce)
math_library(gru_compute DEPS activation_functions math_function) math_library(gru_compute DEPS activation_functions math_function)
math_library(lstm_compute DEPS activation_functions) math_library(lstm_compute DEPS activation_functions)
math_library(math_function DEPS blas dense_tensor tensor) math_library(math_function DEPS blas dense_tensor tensor)
......
...@@ -23,11 +23,12 @@ ...@@ -23,11 +23,12 @@
* @date 2018/6/12 * @date 2018/6/12
**/ **/
#include "paddle/fluid/operators/detection/gpc.h" #include "paddle/phi/kernels/funcs/gpc.h"
#include "paddle/fluid/platform/enforce.h" #include "paddle/phi/core/enforce.h"
namespace gpc { namespace phi {
namespace funcs {
typedef struct lmt_shape { /* Local minima table */ typedef struct lmt_shape { /* Local minima table */
double y; /* Y coordinate at local minimum */ double y; /* Y coordinate at local minimum */
...@@ -541,9 +542,8 @@ static int count_contours(polygon_node *polygon) { ...@@ -541,9 +542,8 @@ static int count_contours(polygon_node *polygon) {
} }
static void add_left(polygon_node *p, double x, double y) { static void add_left(polygon_node *p, double x, double y) {
PADDLE_ENFORCE_NOT_NULL(p, PADDLE_ENFORCE_NOT_NULL(
paddle::platform::errors::InvalidArgument( p, phi::errors::InvalidArgument("Input polygon node is nullptr."));
"Input polygon node is nullptr."));
vertex_node *nv = NULL; vertex_node *nv = NULL;
/* Create a new vertex node and set its fields */ /* Create a new vertex node and set its fields */
...@@ -599,9 +599,8 @@ static void add_right(polygon_node *p, double x, double y) { ...@@ -599,9 +599,8 @@ static void add_right(polygon_node *p, double x, double y) {
} }
static void merge_right(polygon_node *p, polygon_node *q, polygon_node *list) { static void merge_right(polygon_node *p, polygon_node *q, polygon_node *list) {
PADDLE_ENFORCE_NOT_NULL(p, PADDLE_ENFORCE_NOT_NULL(
paddle::platform::errors::InvalidArgument( p, phi::errors::InvalidArgument("Input polygon node is nullptr."));
"Input polygon node is nullptr."));
polygon_node *target = NULL; polygon_node *target = NULL;
/* Label contour as external */ /* Label contour as external */
...@@ -681,8 +680,7 @@ void add_vertex(vertex_node **t, double x, double y) { ...@@ -681,8 +680,7 @@ void add_vertex(vertex_node **t, double x, double y) {
void gpc_vertex_create(edge_node *e, int p, int s, double x, double y) { void gpc_vertex_create(edge_node *e, int p, int s, double x, double y) {
PADDLE_ENFORCE_NOT_NULL( PADDLE_ENFORCE_NOT_NULL(
e, e, phi::errors::InvalidArgument("Input edge node is nullptr."));
paddle::platform::errors::InvalidArgument("Input edge node is nullptr."));
add_vertex(&(e->outp[p]->v[s]), x, y); add_vertex(&(e->outp[p]->v[s]), x, y);
e->outp[p]->active++; e->outp[p]->active++;
} }
...@@ -715,9 +713,8 @@ static bbox *create_contour_bboxes(gpc_polygon *p) { ...@@ -715,9 +713,8 @@ static bbox *create_contour_bboxes(gpc_polygon *p) {
gpc_malloc<bbox>(box, gpc_malloc<bbox>(box,
p->num_contours * sizeof(bbox), p->num_contours * sizeof(bbox),
const_cast<char *>("Bounding box creation")); const_cast<char *>("Bounding box creation"));
PADDLE_ENFORCE_NOT_NULL(box, PADDLE_ENFORCE_NOT_NULL(
paddle::platform::errors::ResourceExhausted( box, phi::errors::ResourceExhausted("Failed to malloc box memory."));
"Failed to malloc box memory."));
/* Construct contour bounding boxes */ /* Construct contour bounding boxes */
for (c = 0; c < p->num_contours; c++) { for (c = 0; c < p->num_contours; c++) {
...@@ -882,9 +879,9 @@ void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) { ...@@ -882,9 +879,9 @@ void gpc_add_contour(gpc_polygon *p, gpc_vertex_list *new_contour, int hole) {
gpc_malloc<int>(extended_hole, gpc_malloc<int>(extended_hole,
(p->num_contours + 1) * sizeof(int), (p->num_contours + 1) * sizeof(int),
const_cast<char *>("contour hole addition")); const_cast<char *>("contour hole addition"));
PADDLE_ENFORCE_NOT_NULL(extended_hole, PADDLE_ENFORCE_NOT_NULL(
paddle::platform::errors::ResourceExhausted( extended_hole,
"Failed to malloc extended hole memory.")); phi::errors::ResourceExhausted("Failed to malloc extended hole memory."));
/* Create an extended contour array */ /* Create an extended contour array */
gpc_malloc<gpc_vertex_list>(extended_contour, gpc_malloc<gpc_vertex_list>(extended_contour,
...@@ -1005,7 +1002,7 @@ void gpc_polygon_clip(gpc_op op, ...@@ -1005,7 +1002,7 @@ void gpc_polygon_clip(gpc_op op,
gpc_malloc<double>( gpc_malloc<double>(
sbt, sbt_entries * sizeof(double), const_cast<char *>("sbt creation")); sbt, sbt_entries * sizeof(double), const_cast<char *>("sbt creation"));
PADDLE_ENFORCE_NOT_NULL(sbt, PADDLE_ENFORCE_NOT_NULL(sbt,
paddle::platform::errors::ResourceExhausted( phi::errors::ResourceExhausted(
"Failed to malloc scanbeam table memory.")); "Failed to malloc scanbeam table memory."));
build_sbt(&scanbeam, sbt, sbtree); build_sbt(&scanbeam, sbt, sbtree);
...@@ -1050,8 +1047,7 @@ void gpc_polygon_clip(gpc_op op, ...@@ -1050,8 +1047,7 @@ void gpc_polygon_clip(gpc_op op,
e1 = aet; e1 = aet;
/* Set up bundle fields of first edge */ /* Set up bundle fields of first edge */
PADDLE_ENFORCE_NOT_NULL( PADDLE_ENFORCE_NOT_NULL(
aet, aet, phi::errors::InvalidArgument("Edge node AET is nullptr."));
paddle::platform::errors::InvalidArgument("Edge node AET is nullptr."));
aet->bundle[ABOVE][aet->type] = (aet->top.y != yb); aet->bundle[ABOVE][aet->type] = (aet->top.y != yb);
aet->bundle[ABOVE][!aet->type] = 0; aet->bundle[ABOVE][!aet->type] = 0;
...@@ -1651,7 +1647,7 @@ void gpc_tristrip_clip(gpc_op op, ...@@ -1651,7 +1647,7 @@ void gpc_tristrip_clip(gpc_op op,
gpc_malloc<double>( gpc_malloc<double>(
sbt, sbt_entries * sizeof(double), const_cast<char *>("sbt creation")); sbt, sbt_entries * sizeof(double), const_cast<char *>("sbt creation"));
PADDLE_ENFORCE_NOT_NULL(sbt, PADDLE_ENFORCE_NOT_NULL(sbt,
paddle::platform::errors::ResourceExhausted( phi::errors::ResourceExhausted(
"Failed to malloc scanbeam table memory.")); "Failed to malloc scanbeam table memory."));
build_sbt(&scanbeam, sbt, sbtree); build_sbt(&scanbeam, sbt, sbtree);
scanbeam = 0; scanbeam = 0;
...@@ -1691,8 +1687,7 @@ void gpc_tristrip_clip(gpc_op op, ...@@ -1691,8 +1687,7 @@ void gpc_tristrip_clip(gpc_op op,
/* Set up bundle fields of first edge */ /* Set up bundle fields of first edge */
PADDLE_ENFORCE_NOT_NULL( PADDLE_ENFORCE_NOT_NULL(
aet, aet, phi::errors::InvalidArgument("Edge node AET is nullptr."));
paddle::platform::errors::InvalidArgument("Edge node AET is nullptr."));
aet->bundle[ABOVE][aet->type] = (aet->top.y != yb); aet->bundle[ABOVE][aet->type] = (aet->top.y != yb);
aet->bundle[ABOVE][!aet->type] = 0; aet->bundle[ABOVE][!aet->type] = 0;
aet->bstate[ABOVE] = UNBUNDLED; aet->bstate[ABOVE] = UNBUNDLED;
...@@ -2248,6 +2243,7 @@ void gpc_tristrip_clip(gpc_op op, ...@@ -2248,6 +2243,7 @@ void gpc_tristrip_clip(gpc_op op,
gpc_free<double>(sbt); gpc_free<double>(sbt);
} // NOLINT } // NOLINT
} // namespace gpc } // namespace funcs
} // namespace phi
/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */ /* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */
...@@ -29,15 +29,16 @@ ...@@ -29,15 +29,16 @@
* @date 2018/6/12 * @date 2018/6/12
**/ **/
#ifndef PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ // GPC_H_ #ifndef PADDLE_PHI_KERNELS_FUNCS_GPC_H_ // GPC_H_
#define PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ // GPC_H_ #define PADDLE_PHI_KERNELS_FUNCS_GPC_H_ // GPC_H_
#include <float.h> #include <float.h>
#include <math.h> #include <math.h>
#include <stdio.h> #include <stdio.h>
#include <stdlib.h> #include <stdlib.h>
namespace gpc { namespace phi {
namespace funcs {
typedef enum { // Set operation type typedef enum { // Set operation type
GPC_DIFF, // Difference GPC_DIFF, // Difference
...@@ -190,7 +191,7 @@ inline void gpc_n_edge(edge_node *d, edge_node *e, int p) { ...@@ -190,7 +191,7 @@ inline void gpc_n_edge(edge_node *d, edge_node *e, int p) {
template <typename T> template <typename T>
void gpc_malloc(T *&p, int b, char *s) { void gpc_malloc(T *&p, int b, char *s) {
if (b > 0) { if (b > 0) {
p = (T *)malloc(b); p = reinterpret_cast<T *>(malloc(b));
if (!p) { if (!p) {
fprintf(stderr, "gpc malloc failure: %s\n", s); fprintf(stderr, "gpc malloc failure: %s\n", s);
...@@ -243,7 +244,8 @@ void gpc_free_polygon(gpc_polygon *polygon); ...@@ -243,7 +244,8 @@ void gpc_free_polygon(gpc_polygon *polygon);
void gpc_free_tristrip(gpc_tristrip *tristrip); void gpc_free_tristrip(gpc_tristrip *tristrip);
} // namespace gpc } // namespace funcs
} // namespace phi
#endif // PADDLE_FLUID_OPERATORS_DETECTION_GPC_H_ #endif // PADDLE_PHI_KERNELS_FUNCS_GPC_H_
/* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */ /* vim: set expandtab ts=4 sw=4 sts=4 tw=100: */
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#pragma once
#include "paddle/phi/core/dense_tensor.h"
namespace phi {
template <typename T, typename Context>
void MultiClassNMSKernel(const Context& ctx,
const DenseTensor& bboxes,
const DenseTensor& scores,
const paddle::optional<DenseTensor>& rois_num,
float score_threshold,
int nms_top_k,
int keep_top_k,
float nms_threshold,
bool normalized,
float nms_eta,
int background_label,
DenseTensor* out,
DenseTensor* index,
DenseTensor* nms_rois_num);
} // namespace phi
// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
//
// Licensed under the Apache License, Version 2.0 (the "License");
// you may not use this file except in compliance with the License.
// You may obtain a copy of the License at
//
// http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing, software
// distributed under the License is distributed on an "AS IS" BASIS,
// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
// See the License for the specific language governing permissions and
// limitations under the License.
#include "paddle/phi/core/compat/op_utils.h"
namespace phi {
KernelSignature MultiClassNMS3OpArgumentMapping(
const ArgumentMappingContext& ctx) {
return KernelSignature("multiclass_nms3",
{"BBoxes", "Scores", "RoisNum"},
{"score_threshold",
"nms_top_k",
"keep_top_k",
"nms_threshold",
"normalized",
"nms_eta",
"background_label"},
{"Out", "Index", "NmsRoisNum"});
}
} // namespace phi
PD_REGISTER_ARG_MAPPING_FN(multiclass_nms3,
phi::MultiClassNMS3OpArgumentMapping);
...@@ -1457,6 +1457,7 @@ class OpTest(unittest.TestCase): ...@@ -1457,6 +1457,7 @@ class OpTest(unittest.TestCase):
# see details: https://stackoverflow.com/questions/38331703/why-does-numpys-broadcasting-sometimes-allow-comparing-arrays-of-different-leng # see details: https://stackoverflow.com/questions/38331703/why-does-numpys-broadcasting-sometimes-allow-comparing-arrays-of-different-leng
if expect_np.size == 0: if expect_np.size == 0:
self.op_test.assertTrue(actual_np.size == 0) # }}} self.op_test.assertTrue(actual_np.size == 0) # }}}
# print("actual_np, expect_np", actual_np, expect_np)
self._compare_numpy(name, actual_np, expect_np) self._compare_numpy(name, actual_np, expect_np)
if isinstance(expect, tuple): if isinstance(expect, tuple):
self._compare_list(name, actual, expect) self._compare_list(name, actual, expect)
......
...@@ -19,7 +19,81 @@ import copy ...@@ -19,7 +19,81 @@ import copy
from op_test import OpTest from op_test import OpTest
import paddle import paddle
import paddle.fluid as fluid import paddle.fluid as fluid
from paddle.fluid import Program, program_guard from paddle.fluid import Program, program_guard, in_dygraph_mode, _non_static_mode
from paddle.fluid.layer_helper import LayerHelper
from paddle import _C_ops
def multiclass_nms3(bboxes,
scores,
rois_num=None,
score_threshold=0.3,
nms_top_k=1000,
keep_top_k=100,
nms_threshold=0.3,
normalized=True,
nms_eta=1.,
background_label=-1,
return_index=True,
return_rois_num=True,
name=None):
helper = LayerHelper('multiclass_nms3', **locals())
if in_dygraph_mode():
attrs = (score_threshold, nms_top_k, keep_top_k, nms_threshold,
normalized, nms_eta, background_label)
output, index, nms_rois_num = _C_ops.final_state_multiclass_nms3(
bboxes, scores, rois_num, *attrs)
if not return_index:
index = None
return output, index, nms_rois_num
elif _non_static_mode():
attrs = ('background_label', background_label, 'score_threshold',
score_threshold, 'nms_top_k', nms_top_k, 'nms_threshold',
nms_threshold, 'keep_top_k', keep_top_k, 'nms_eta', nms_eta,
'normalized', normalized)
output, index, nms_rois_num = _C_ops.multiclass_nms3(
bboxes, scores, rois_num, *attrs)
if not return_index:
index = None
return output, index, nms_rois_num
else:
output = helper.create_variable_for_type_inference(dtype=bboxes.dtype)
index = helper.create_variable_for_type_inference(dtype='int32')
inputs = {'BBoxes': bboxes, 'Scores': scores}
outputs = {'Out': output, 'Index': index}
if rois_num is not None:
inputs['RoisNum'] = rois_num
if return_rois_num:
nms_rois_num = helper.create_variable_for_type_inference(
dtype='int32')
outputs['NmsRoisNum'] = nms_rois_num
helper.append_op(type="multiclass_nms3",
inputs=inputs,
attrs={
'background_label': background_label,
'score_threshold': score_threshold,
'nms_top_k': nms_top_k,
'nms_threshold': nms_threshold,
'keep_top_k': keep_top_k,
'nms_eta': nms_eta,
'normalized': normalized
},
outputs=outputs)
output.stop_gradient = True
index.stop_gradient = True
if not return_index:
index = None
if not return_rois_num:
nms_rois_num = None
return output, nms_rois_num, index
def softmax(x): def softmax(x):
...@@ -541,8 +615,9 @@ class TestMulticlassNMS2LoDInput(TestMulticlassNMSLoDInput): ...@@ -541,8 +615,9 @@ class TestMulticlassNMS2LoDInput(TestMulticlassNMSLoDInput):
'normalized': normalized, 'normalized': normalized,
} }
def test_check_output(self):
self.check_output() def test_check_output(self):
self.check_output()
class TestMulticlassNMS2LoDNoOutput(TestMulticlassNMS2LoDInput): class TestMulticlassNMS2LoDNoOutput(TestMulticlassNMS2LoDInput):
...@@ -590,6 +665,7 @@ class TestMulticlassNMSError(unittest.TestCase): ...@@ -590,6 +665,7 @@ class TestMulticlassNMSError(unittest.TestCase):
class TestMulticlassNMS3Op(TestMulticlassNMS2Op): class TestMulticlassNMS3Op(TestMulticlassNMS2Op):
def setUp(self): def setUp(self):
self.python_api = multiclass_nms3
self.set_argument() self.set_argument()
N = 7 N = 7
M = 1200 M = 1200
...@@ -623,8 +699,8 @@ class TestMulticlassNMS3Op(TestMulticlassNMS2Op): ...@@ -623,8 +699,8 @@ class TestMulticlassNMS3Op(TestMulticlassNMS2Op):
self.op_type = 'multiclass_nms3' self.op_type = 'multiclass_nms3'
self.inputs = {'BBoxes': boxes, 'Scores': scores} self.inputs = {'BBoxes': boxes, 'Scores': scores}
self.outputs = { self.outputs = {
'Out': (nmsed_outs, [lod]), 'Out': nmsed_outs,
'Index': (index_outs, [lod]), 'Index': index_outs,
'NmsRoisNum': np.array(lod).astype('int32') 'NmsRoisNum': np.array(lod).astype('int32')
} }
self.attrs = { self.attrs = {
...@@ -638,7 +714,7 @@ class TestMulticlassNMS3Op(TestMulticlassNMS2Op): ...@@ -638,7 +714,7 @@ class TestMulticlassNMS3Op(TestMulticlassNMS2Op):
} }
def test_check_output(self): def test_check_output(self):
self.check_output() self.check_output(check_eager=True)
class TestMulticlassNMS3OpNoOutput(TestMulticlassNMS3Op): class TestMulticlassNMS3OpNoOutput(TestMulticlassNMS3Op):
...@@ -649,71 +725,6 @@ class TestMulticlassNMS3OpNoOutput(TestMulticlassNMS3Op): ...@@ -649,71 +725,6 @@ class TestMulticlassNMS3OpNoOutput(TestMulticlassNMS3Op):
self.score_threshold = 2.0 self.score_threshold = 2.0
class TestMulticlassNMS3LoDInput(TestMulticlassNMS2LoDInput):
def setUp(self):
self.set_argument()
M = 1200
C = 21
BOX_SIZE = 4
box_lod = [[1200]]
background = 0
nms_threshold = 0.3
nms_top_k = 400
keep_top_k = 200
score_threshold = self.score_threshold
normalized = False
scores = np.random.random((M, C)).astype('float32')
scores = np.apply_along_axis(softmax, 1, scores)
boxes = np.random.random((M, C, BOX_SIZE)).astype('float32')
boxes[:, :, 0] = boxes[:, :, 0] * 10
boxes[:, :, 1] = boxes[:, :, 1] * 10
boxes[:, :, 2] = boxes[:, :, 2] * 10 + 10
boxes[:, :, 3] = boxes[:, :, 3] * 10 + 10
det_outs, lod = lod_multiclass_nms(boxes, scores, background,
score_threshold, nms_threshold,
nms_top_k, keep_top_k, box_lod,
normalized)
det_outs = np.array(det_outs)
nmsed_outs = det_outs[:, :-1].astype('float32') if len(
det_outs) else det_outs
self.op_type = 'multiclass_nms3'
self.inputs = {
'BBoxes': (boxes, box_lod),
'Scores': (scores, box_lod),
'RoisNum': np.array(box_lod).astype('int32')
}
self.outputs = {
'Out': (nmsed_outs, [lod]),
'NmsRoisNum': np.array(lod).astype('int32')
}
self.attrs = {
'background_label': 0,
'nms_threshold': nms_threshold,
'nms_top_k': nms_top_k,
'keep_top_k': keep_top_k,
'score_threshold': score_threshold,
'nms_eta': 1.0,
'normalized': normalized,
}
def test_check_output(self):
self.check_output()
class TestMulticlassNMS3LoDNoOutput(TestMulticlassNMS3LoDInput):
def set_argument(self):
# Here set 2.0 to test the case there is no outputs.
# In practical use, 0.0 < score_threshold < 1.0
self.score_threshold = 2.0
if __name__ == '__main__': if __name__ == '__main__':
paddle.enable_static() paddle.enable_static()
unittest.main() unittest.main()
Markdown is supported
0% .
You are about to add 0 people to the discussion. Proceed with caution.
先完成此消息的编辑!
想要评论请 注册